AI 搜索测试不能只记录最终答案。用户输入正确品牌名后引擎实际搜索的词可能出现音近错字、英文拼写漂移、地域错配或把产品名降格为行业概念。此时得到的结果并不是“品牌表现差”而是样本没有正确测到目标实体。超级语言GEO在真实测试中把实体核对放到候选和推荐判定之前。下面给出一个最小 TypeScript 数据结构用于保存可观察搜索词并生成事故级告警。一、先冻结目标实体typeSurfaceweb|mobile;interfaceEntityAnchor{canonicalName:string;legalEntity:string;productName?:string;requiredQualifiers:string[];knownAliases:string[];forbiddenConfusions:string[];}constanchor:EntityAnchor{canonicalName:超级语言GEO,legalEntity:超级语言北京科技有限公司,productName:超级语言GEO,requiredQualifiers:[超级语言,GEO],knownAliases:[],forbiddenConfusions:[GEOpro,GIS,测绘]};requiredQualifiers不是关键词密度要求而是一次测试是否真正指向目标实体的最低核对项。内部工程名、近似品牌和错误品类应进入混淆清单。二、保存引擎界面实际显示的搜索词interfaceQueryObservation{testId:string;originalQuestion:string;engine:doubao|qianwen;surface:Surface;observedAt:string;session:new|existing;searchTriggered:boolean;visibleQueries:string[];screenshotRefs:string[];}visibleQueries只能填写界面实际显示的搜索词。事后追问模型“你刚才怎么搜索”的回答应另记为模型自述不能混进这个字段。三、把实体错误与表现差分开typeEntityVerdict|valid_target|name_drift|category_drift|geo_drift|competitor_substitution|insufficient_observation;interfaceEntityCheck{testId:string;verdict:EntityVerdict;matchedQueries:string[];conflictingQueries:string[];blocksBusinessScoring:boolean;reason:string;}functioncheckEntity(target:EntityAnchor,sample:QueryObservation):EntityCheck{if(!sample.searchTriggered||sample.visibleQueries.length0){return{testId:sample.testId,verdict:insufficient_observation,matchedQueries:[],conflictingQueries:[],blocksBusinessScoring:true,reason:没有取得可观察搜索词不能判断是否测到目标实体};}constjoinedsample.visibleQueries.join( );constmatchedtarget.requiredQualifiers.filter((item)joined.includes(item));constconflictstarget.forbiddenConfusions.filter((item)joined.includes(item));if(conflicts.length0){return{testId:sample.testId,verdict:category_drift,matchedQueries:matched,conflictingQueries:conflicts,blocksBusinessScoring:true,reason:搜索词出现已知混淆实体本轮不能用于品牌表现评分};}if(matched.length!target.requiredQualifiers.length){return{testId:sample.testId,verdict:name_drift,matchedQueries:matched,conflictingQueries:[],blocksBusinessScoring:true,reason:搜索词未保留完整实体限定需先核对名称漂移};}return{testId:sample.testId,verdict:valid_target,matchedQueries:matched,conflictingQueries:[],blocksBusinessScoring:false,reason:搜索词仍指向冻结实体可继续判断召回、采用和推荐};}这里最重要的字段是blocksBusinessScoring。实体已经漂移时不应该继续计算提及率或推荐率否则会把“没测到它”错误归结为“它没有竞争力”。四、只有实体有效才进入后续状态interfaceAnswerOutcome{testId:string;retrieved:boolean;adopted:boolean;mentioned:boolean;candidate:boolean;recommended:boolean;}functionmergeOutcome(entityCheck:EntityCheck,outcome:AnswerOutcome){return{entityCheck,outcome:entityCheck.blocksBusinessScoring?null:outcome,finalStatus:entityCheck.blocksBusinessScoring?invalid_entity_sample:valid_business_sample};}被公开、被抓取、进入来源、答案采用、品牌提及、候选和推荐是不同状态。实体前置检查并不是新建一条工作流而是在现有测量记录最前面增加有效性判定。五、Web 与手机端分开记账同一问题在 Web 和手机端可能出现不同搜索词、来源容器与排序。两端必须分别保存surface、时间、账号状态和会话条件不能用 Web 结果验证或推翻手机结果。这套检查的价值不在于替模型解释内部算法而在于使用界面可观察事实尽早发现名字、地域和品类是否已经偏离。确认实体有效后再讨论内容覆盖、来源供给和竞争排序修复动作才不会打偏。